Recuperação de poses humanas 3D a partir de imagens bidimensionais

Abstract

This thesis presents a new model to identify 3D human poses from photos. Given a single input image and a new model for characterization of postures based on their "comfort", this work aims to solve ambiguous postures and to generate the possible postures in 3D. The 3D human pose recovery model is based on the method of Taylor [Tay00] and includes improvements with biomechanical constraints in order to reduce the search space of possible 3D postures that can represent the pose in the 2D image. Furthermore, it is proposed a classification system which can be used to suggest the best positions generated according to the features of comfort and lightness of the image which are also used. The comfort criteria is related to assumptions in terms of the posture balance, while the shade criterion eliminates the ambiguities of postures, taking into account the illumination in the image. It should be emphasized that the recovery poses in 3D from just a single two-dimensional image is the main focus of this work. The paper also proposes one application for minimizing ambiguous poses based on self-occlusion. Result analysis was performed to check the validity of the models proposed in this thesis.